NextFin News — Chinese e-commerce and delivery giant Meituan announced Monday that it has officially open-sourced its next-generation, trillion-parameter artificial intelligence model, LongCat-2.0, along with its dedicated domestic inference source code.
The sparse Mixture-of-Experts (MoE) architecture features a massive 1.6-trillion total parameter footprint while constraining active runtime computation to an average of 48 billion parameters per token. Engineered specifically to bypass the memory bandwidth and hardware limitations of domestic graphics processing units and accelerators, the system natively supports a 1-million-token context window through integrated structural co-optimizations across its model layers, chip adaptation protocols, and deployment strategies.
Pressure is mounting on technology conglomerates on the Chinese mainland to achieve operational independence from international silicon suppliers by migrating heavy computing workloads to domestic hardware clusters. By open-sourcing a multi-trillion-parameter model optimized from inception to run efficiently on alternative domestic architecture, Meituan is actively lowering the deployment friction and infrastructure costs typically associated with running large-scale generative software. For corporate enterprise clients and cloud infrastructure providers, this strategic release accelerates the democratization of near-frontier agentic programming and data processing pipelines, shifting the industry standard away from rigid hardware dependency toward software-driven architectural efficiency.
Explore more exclusive insights at nextfin.ai.
